SambaNova
SambaNova Systems, founded in 2017 by Stanford professors Kunle Olukotun and Christopher Ré alongside ex-Oracle executive Rodrigo Liang, builds custom processors and software for AI inference at enterprise scale.
Profile
Custom AI inference chips and software for enterprise and sovereign AI deployments.
SambaNova Systems, founded in 2017 by Stanford professors Kunle Olukotun and Christopher Ré alongside ex-Oracle executive Rodrigo Liang, builds custom processors and software for AI inference at enterprise scale. The company architected its RDU (Reconfigurable Dataflow Unit) processor to accelerate token generation for large language models, positioning itself as an alternative to Nvidia-dominated AI infrastructure. By 2021, SambaNova had raised a $676 million Series D and reached a $5.1 billion valuation, backed by SoftBank, Google Ventures, and other leading investors.
The company sold cloud-based inference through SambaCloud and hardware as a solution to GPU shortages and power consumption constraints.\n\nThat positioning fractured in 2024–2025 as GPU supply normalized and the AI market shifted from training to inference-heavy workloads. In April 2025, SambaNova laid off 77 employees (15% of its 500-person workforce) and abandoned its training cloud business entirely, refocusing exclusively on inference. The move signaled material traction challenges—the company had oversold its relevance to the training economy just as that market matured.
Meanwhile, acquisition talks with Intel stalled after the company rejected a proposed $1.6 billion buyout.\n\nIn February 2026, SambaNova closed a $350 million Series E round led by Vista Equity Partners and Cambium Capital, with Intel as a co-investor, at an implied valuation of approximately $2.2 billion—a 57% decline from the 2021 peak. The round reflected investor recalibration around non-Nvidia AI chips.\n\nThe company's new strategy centers on SambaManaged, a turnkey inference platform deployable in 90 days (vs. 18–24 months typical), and the new SN50 chip, which targets agentic AI workloads and claims 3x power efficiency over competing inference processors. In October 2025, SambaNova announced partnerships with sovereign AI providers in Australia, Europe (Infercom), and the UK—positioning itself as the inference backbone for regionally-compliant AI clouds.
Named customers include Los Alamos National Laboratory, Argonne National Laboratory, Hugging Face, Meta, and SoftBank, which committed to deploying the SN50 in Japanese data centers. The company expects to begin H2 2026 enterprise deployments.
Who buys this
- Government and national labs (Los Alamos, Argonne) running AI inference at scale
- Sovereign AI cloud providers in Europe, Australia, and UK requiring data residency compliance
- Enterprise ML teams building private or fine-tuned LLM deployments
- Inference-as-a-Service platforms and cloud providers (Hugging Face)
- Large tech and hyperscaler companies requiring energy-efficient token generation (Meta, SoftBank)
Publicly disclosed clients
- Los Alamos National Laboratory
- Argonne National Laboratory
- Hugging Face
- Meta
- SoftBank
Strengths and what to watch
Strengths
- Purpose-built RDU/SN50 inference chip with demonstrable power efficiency (10kW per rack vs. 120kW for GPU systems), eliminating liquid-cooling cost and complexity
- SambaManaged platform deployable in 90 days vs. typical 18–24 month data center buildouts; proven adoption by Infercom for sovereign AI cloud launch
- Beachhead in government and sovereign AI markets (US DoE labs, Infercom/Germany, Argyll/UK, SCX/Australia) where compliance and data residency create switching costs
Watch for
- Valuation collapsed 57% from $5.1B (April 2021) to ~$2.2B (February 2026 Series E); investor appetite for non-Nvidia AI infrastructure remains unproven at scale
- April 2025 layoffs (15% workforce reduction) followed failed pivot from training to inference, signaling execution risk; commercial traction outside government/sovereign cloud not yet visible
- Customer concentration in government and regional sovereign AI providers; enterprise commercial pipeline and SoftBank deployment timeline (SN50 availability pegged H2 2026) are key 2026 watch points
Recent moves
Key Information
- Industry
- AI Hardware
- Founded
- 2017
- Employees
- 400
- Headquarters
- United States
Frequently Asked Questions
What does SambaNova do?
SambaNova Systems builds custom AI inference chips and software for enterprises and sovereign AI clouds. Founded in 2017 by Stanford professors and an ex-Oracle executive, the company designs the RDU processor to accelerate token generation as a non-Nvidia alternative for AI deployments.
What is the SN50 chip?
The SN50 is SambaNova's latest inference processor targeting agentic AI workloads. It claims 3x power efficiency over competing inference processors and operates at 10kW per rack versus 120kW for typical GPU systems, eliminating expensive liquid-cooling costs and infrastructure complexity for enterprises.
How long does SambaManaged take to deploy?
SambaManaged is SambaNova's turnkey inference platform deployable in 90 days, versus 18–24 months required for typical data center buildouts. Infercom, a German sovereign AI provider, adopted the platform to rapidly launch a regional, compliance-focused AI cloud service, validating the accelerated deployment model.
Who uses SambaNova?
SambaNova serves government labs like Los Alamos and Argonne, sovereign AI providers in Europe, Australia, and UK requiring data residency compliance, and major tech companies including Meta and SoftBank. The company is expanding into enterprise ML teams and inference-as-a-service platforms globally.
Why did SambaNova's valuation drop in 2025?
SambaNova's valuation fell 57% from $5.1 billion in 2021 to $2.2 billion in February 2026 after the company laid off 15% of its workforce and pivoted away from training workloads to focus on inference. GPU supply normalization reduced demand for alternatives.
How does SambaNova compete with Nvidia?
SambaNova targets inference workloads where custom chips offer power efficiency gains unavailable with GPUs. It competes on deployment speed (90 days vs. months) and regulatory compliance for sovereign AI clouds. However, Nvidia's dominance in training and mature ecosystem remain advantages.
How SambaNova compares
Direct head-to-head against 2 competitors. Picked by 7wData.
SambaNova
- Positioning
- Custom AI inference chips and software for enterprise and sovereign AI deployments.
- Customer segments
- Government and national labs (Los Alamos, Argonne) running AI inference at scale
- Strengths
- Purpose-built RDU/SN50 inference chip with demonstrable power efficiency (10kW per rack vs. 120kW for GPU systems), eliminating liquid-cooling cost and complexity
- Watch for
- Valuation collapsed 57% from $5.1B (April 2021) to ~$2.2B (February 2026 Series E); investor appetite for non-Nvidia AI infrastructure remains unproven at scale
- Recent moves
- SambaNova raises $350 million Series E at $2.2B valuation, announces SN50 chip and Intel partnership
Cerebras Systems
- Positioning
- Custom wafer-scale AI chip company targeting inference speed; IPO filed April 2026, valued near $23 billion.
- Customer segments
- Government-linked AI institutions (UAE, Saudi Arabia), enterprise ML teams, cloud providers; buyer persona is CTO-level infrastructure.
- Strengths
- Wafer-Scale Engine chip holds highest single-chip memory bandwidth available; benchmarked fastest inference platform by Artificial Analysis in 2025.
- Watch for
- 86% of 2025 revenue tied to two UAE entities (G42, MBZUAI); CFIUS review blocked first IPO attempt in late 2024.
- Recent moves
- Multi-year AWS distribution partnership announced March 2026; IPO S-1 refiled April 2026 after CFIUS clearance in October 2025.
Groq
- Positioning
- LPU inference chip company operating independently after December 2025 Nvidia licensing deal; cloud-first platform for developer and enterprise inference.
- Customer segments
- Developers (nearly 2M on GroqCloud), enterprise AI teams, sovereign AI funds; buyer persona is ML engineer and infrastructure lead.
- Strengths
- LPU architecture delivers deterministic, low-latency token generation; secured $1.5B Saudi Arabia commitment for 19,000-unit inference cluster in February 2025.
- Watch for
- Founder Jonathan Ross and president Sunny Madra departed to Nvidia in December 2025; independent product roadmap under new CEO Simon Edwards is unproven.
- Recent moves
- Nvidia licensed Groq LPU technology for approximately $20 billion in December 2025; Ross and Madra joined Nvidia; Edwards appointed CEO.
Sources
- sambanova.ai — Company mission, products (SambaCloud, SambaStack, SambaManaged), RDU architecture, and customer partnerships
- sambanova.ai — Founder names and backgrounds (Kunle Olukotun, Christopher Ré, Rodrigo Liang), founding year (2017), company mission and history
- sambanova.ai — Executive leadership team (CEO Rodrigo Liang, Chief Technologist Kunle Olukotun, CPSO Abhi Ingle, CRO Harry Ault, others)
- www.crunchbase.com — Total funding history (~$1.48B across 5 rounds as of April 2026), Series E details, investor list (Vista Equity, Intel, SoftBank, Google Ventures, BlackRock, etc.)
- www.bloomberg.com — Series E funding round ($350M, February 2026), valuation ($2.2B), investor participation (Vista, Intel, others), SN50 announcement, SoftBank deployment commitment
- www.businesswire.com — October 2025 sovereign AI partnerships (Infercom/Germany, Argyll/UK, SCX/Australia), SambaManaged 90-day deployment claim
- www.datacenterdynamics.com — April 22, 2025 layoff details (77 employees, 15% of ~500 workforce), strategic pivot from training to inference
- www.alcf.anl.gov — Argonne National Laboratory customer verification and deployment details for AI inference workloads